Research On Link Prediction Method Of Knowledge Graph Based On Neural Network

Text Knowledge graphs are institutionalized representations of real-world information.In a knowledge graph,entities are represented by nodes,and edges are specific relationships connecting two entities.Because of its ability to model structured data in a computer-readable way,knowledge graphs are widely used in natural language processing,such as intelligent question and answer and content-based recommendation systems.Despite the richness of the data contained in the knowledge graph,the completeness of the knowledge graph currently leaves much to be desired,leading researchers to investigate how to identify new knowledge to complement the knowledge graph,a task often referred to as the knowledge graph complementation task or the knowledge graph expansion task.Of course,we can make inferences about new knowledge based on the knowledge that already exists in the knowledge graph,a task also known as link prediction(link prediction),and the link prediction task is the focus of our analysis.The current knowledge graph link prediction model suffers from a number of drawbacks:(1)Some neural network-based models make use of multiple convolutional layers in order to improve the performance,which makes the model structure deeper.Due to the excessive number of convolutional layers,it will have a negative effect on the preservation of surface semantic knowledge,which will lead to the loss of surface knowledge to some extent.(2)Most convolutional neural network models only consider combining the features of head entities and relations,without considering the correlation of triads in the same dimension,ignoring the overall features of triads and the translation characteristics.From the above problems,the main research of this thesis is as follows:(1)For the first problem,in order to retain explicit knowledge and enhance the extraction capability for potential knowledge,this thesis gives proposes a knowledge graph link prediction model with combined 1D and 2D convolutional embeddings.In this thesis,vector multiplication operation is performed on entity embedding and relationship embedding to generate a 2D matrix that facilitates 2D convolution,and each element in the matrix is the product of the corresponding element in entity embedding and relationship embedding to achieve element-level fusion,and this mechanism greatly facilitates the interaction between entities and relationships.Also 1D convolution is appropriately used to perform convolution operations on the input entity embeddings and relational embeddings to extract surface and explicit knowledge.(2)For the second problem,a knowledge graph link prediction model based on relational memory network and convolutional neural network is proposed in this thesis.In this thesis,the input sequences of head entities,relations and tail entities are encoded with positions.Then,this thesis uses Transformer self-attentive mechanism to interact with the memory matrix to generate the encoding vector.It can effectively reason about the complex semantic relationships between entities and relations and capture the deep relationships between entity and relation embedding vectors.Meanwhile,in the convolutional decoder part,this study proposes a matrix shaping strategy,which greatly increases the feature interaction between entities and relations in more dimensions.(3)In this thesis,experimental analysis of link prediction is carried out on several data sets,and the two models proposed in this thesis outperform the baseline model in most metrics,which proves the practicality of the algorithms in this thesis.